Adrian Seyboldt

Just wanted to ask if there is anything new about this? I think this change would still be quite useful.

Do you mean ``` @numba.njit def inc_array(x, val): x[0] += val[()] ``` because otherwise this will increment all locations in x, not just the first.

Looks good. I think it is possible that we could further improve the MvNormal in both parametrizations, but this is definetly an improvement as it is. Most of all I...

Thanks for reporting. I can't find any documentation about what API they have to show output, and without that, it's pretty hard to fix this. Do you have any idea...

I think it would be great if we could improve this! But if we change names, maybe we could also change the design a little bit at the same time?...

> Are back_prop, fwd_prop very wrong? I've never seen those before... I guess I like pullback and pushforward, because those are simply the names of those functions in math. :-)...

Where would something like ExpAndGrad ever come from? ``` class Exp(Op): def push_forward(self, primals, tangents): (x,) = primals (d_x,) = tangents y = self(x) return y, [y * d_x] #...

The problem with the current way is that when we compute the tangents of the output we sometimes have to generate a second way to compute the output (see the...

So for the loop it might look something like this: ```python class Loop(Op): def push_forward(self, node, tangents, need_output_tangent: List[bool]): push_forward_update = push_forward(self.update, ...not for the go_on arg...) # not implement...

The nomenclature is certainly a bit messy at the moment about this whole thing. I don't think "pushforward" and "pullback" are that common in the autodiff literature right now, but...